What is the difference between the OpenAI Agents API and the Agents SDK?

The Agents SDK is an open-source library that runs the agent loop inside your own process — you host it, you own the tool execution, you decide how to compact context. The Agents API, launched in public beta on September 10, 2026, is a hosted loop: OpenAI runs the Codex harness, provisions a sandbox (OpenAI-hosted, self-hosted or a partner such as E2B, Modal or Cloudflare), compacts context, delegates to subagents and persists the session. OpenAI's docs are explicit that an Agents API session, an SDK session and a Responses conversation are different resources — you cannot swap one for another without a migration.

Should I use the Responses API or the Agents API for an AI agent in 2026?

Use the Responses API when you want a stateless model call you fully control, need Zero Data Retention, need non-US data residency, or are building a short interactive loop. Use the Agents API when the agent must run for hours or days, needs a sandbox with files and code execution, benefits from parallel subagents, or you simply do not want to maintain compaction and tool-routing code. The Agents API adds no fee — you pay the same token rates plus tool and container charges — but it currently offers US-only residency and no ZDR.

Does the OpenAI Agents API cost more than the Agents SDK?

Not in platform fees. As of September 2026 the Agents API charges nothing beyond the model's standard token rates (for example GPT-6 Astra at $10/$50 per million tokens), built-in tools at standard rates, and OpenAI-hosted sandboxes at standard container rates. The SDK is free software, but you pay for your own compute. The realistic cost difference is engineering time: the Agents API includes compaction, tool search, programmatic tool calling and multi-agent orchestration that the SDK expects you to configure.

Can I self-host the sandbox with the OpenAI Agents API?

Yes. Set environment type to self_hosted with a workspace_directory and optional capability_directories, or use one of nine partner sandboxes — Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. The agent's commands and files then stay on your infrastructure. Session state and orchestration still run on OpenAI's side, which is why self-hosting does not make the Agents API ZDR-eligible.

Quick Answer

Agents API vs Agents SDK vs Responses API: OpenAI Compared

Published: September 11, 2026

The Short Answer

	Responses API	Agents SDK	Agents API (new, Sep 10, 2026)
What it is	One model call; you build the loop	Open-source library that runs the loop in your process	Hosted Codex harness that runs the loop for you
Who runs orchestration	You	You (library code)	OpenAI
State	Optional conversation object	In your process/store	Durable session on OpenAI’s side
Sandbox / code execution	Built-in tools only (e.g. code interpreter)	Whatever you wire up	OpenAI-hosted, self-hosted, or 9 partner sandboxes
Compaction	DIY	Configurable in code	Automatic near the context limit
Subagents	DIY	Handoffs/agents-as-tools you define	`multi_agent` with `max_concurrent_subagents`
Long sessions (hours–days)	Painful	Possible with your infra	Designed for it; steer and resume
Extra platform fee	None	None	None — tokens + tools + container rates
ZDR	Available per endpoint policy	Depends on the endpoint you call	Not supported
Data residency	Per platform policy	Per endpoint	US only
Maturity	GA	GA	Public beta

Rule of thumb: Responses for control and compliance, Agents SDK for a loop you own, Agents API for cloud agents that must run long and you’d rather not babysit.

Responses API: the raw material

The Responses API is a single model turn with tools attached. It is the right primitive when the “agent” is a few steps of tool use inside a request you already control, when you need Zero Data Retention or regional data residency, or when latency and predictability beat autonomy. Everything else — deciding when to stop, storing history, summarising context, running commands somewhere — is your code.

Agents SDK: your loop, your servers

The Agents SDK wraps the loop: agents, tools, handoffs, guardrails, tracing. It stays in your process, which means you own tool execution and the data path end-to-end. That is exactly why it remains the answer for teams with strict security review — nothing runs anywhere they did not deploy it. The price is maintenance: every new model capability (tool search, programmatic tool calling, better compaction) is something you adopt and re-test. OpenAI’s own launch note for the Agents API says this out loud — “taking advantage of new model capabilities often means reworking your harness.”

Agents API: OpenAI runs the harness

The Agents API is the Codex harness as a service. One POST /v1/agents/sessions call with the model (gpt-6-astra in the docs), instructions, tools (programmatic_tool_calling, mcp, web_search, …), a multi_agent block and an environment creates a durable session. OpenAI then:

provisions the sandbox (or connects to yours),
compacts context automatically as the session nears its limit,
loads tool definitions on demand and lets the agent filter results in code,
spins up parallel subagents with isolated contexts,
lets you steer mid-turn and resume later,
emits events by stream or webhook.

The harness is versioned with each model launch, so the improvements arrive without you rewriting the loop. And the core logic is public — the open-source Codex repo — so you can read what it does.

Two hard constraints in the docs as of September 11, 2026: US data residency only and no ZDR (self-hosting the sandbox does not change this). If either is a requirement, this is not your API yet.

Cost, honestly

None of the three charges a platform fee. Token spend is identical for the same model and the same amount of work. The differences:

Agents API adds container charges if you use an OpenAI-hosted sandbox; partner sandboxes bill you directly; self-hosted is your compute.
Agents API may reduce tokens through tool search (definitions loaded on demand, cache preserved) and programmatic tool calling (large intermediate results never enter context). Whether this beats a well-tuned SDK loop is workload-specific — measure it on a hundred real tasks.
Engineering time is the real line item. Compaction, subagent coordination and long-session recovery are weeks of work in the SDK and configuration in the Agents API.

Decision guide

Situation	Pick
ZDR contract, EU/UK data residency, regulated data	Responses API (or SDK on a ZDR-eligible endpoint)
Short interactive assistant inside your app	Responses API
You need to own every byte of the execution path	Agents SDK
Multi-hour research, coding or data-analysis agents	Agents API
Many parallel subtasks with a merge step	Agents API (`multi_agent`)
You already run an SDK loop and it works	Stay — the Agents API is a migration, not an upgrade switch
Prototype in an afternoon with a sandbox	Agents API, OpenAI-hosted sandbox

What to watch

The Agents API is in public beta; OpenAI says it will “iterate quickly” toward GA. The two things that would change this comparison most are ZDR support and non-US residency. Until then, the honest split is: compliance-first teams stay on Responses/SDK, velocity-first teams move heavy agents to the Agents API.